yetone
Featured
kill-ai-slop
Kill AI Slop is a multilingual field guide and agent skill that catalogues 33 common AI-generated design tells with interactive before-after demos, and provides a tool to scan and fix web projects.
qingyi-kda
Qingyi-KDA is a 0.6B-parameter hybrid linear-attention language model that successfully converts Qwen3-0.6B-Base into a KDA-based architecture, achieving the smallest publicly released trained KDA model with a stable persona baked into weights.
Stars
7
7-day growth
No data
Forks
1
Open issues
0
License
Apache-2.0
Last updated
2026-07-30
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It is the first successful Qwen3→KDA linearization (previous attempts reported divergence), the smallest sub-1B KDA model with practical language ability, and the first open-source linear-attention model with a fully baked persona, all trained on a single consumer GPU.
Who it is for
- Researchers in efficient transformer architectures and linear attention
- Developers needing small, fast conversational AI with low memory footprint
- Hobbyists exploring model distillation and hybrid attention on consumer hardware
- Open-source community interested in reproducible Qwen3-to-KDA conversion pipelines
Use cases
- Running a lightweight chat assistant on edge devices or low-resource GPUs
- Studying and extending hybrid attention (GQA+KDA) for language modeling
- Prototyping persona-driven chatbots without requiring system prompts
Strengths
- First successful conversion from Qwen3 to KDA, overcoming documented convergence failure
- Smallest publicly available trained KDA model (0.6B) with practical conversational ability
- End-to-end linearization on a single RTX 4070 Laptop 8GB, demonstrating feasibility for consumer hardware
- Persona alignment without system prompt, with privacy extraction audit showing no real-name leakage
Considerations
- 0.6B capacity limits: long responses may repeat, factual reliability is weak, general knowledge is limited
- Trained on group-chat ChatML format; bare QA format is out-of-distribution, degrading performance
- Persona data includes adult content; model may generate NSFW text and requires appropriate usage context
README quick start
快速开始
Description
QINGYI-KDA-0.6B: first successful Qwen3-to-KDA linearization (3:1 hybrid, GQA+RoPE kept layers) + persona alignment. Smallest trained KDA model, converted end-to-end on a consumer 8GB GPU.
Related repositories
Similar projects matched by category, topics, and programming language.
gnipbao
Featured
story-to-handdrawn-video
A Remotion-based tool that converts Chinese story text or ordered hand-drawn images into vertical hand-drawn diary-comic animation with handwritten captions, left-to-right reveals, optional page-curl transitions, and silent H.264 output for post-production dubbing.
lopopolo
Featured
harness-engineering
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.